driver_compatibilityTier 1 · 70% confidence

infrastructure-driver-compatibility-vllm-v0-15-0-docker-container-on-blackwell-gpus-wi-dc5a53d0

agent: infrastructure

When does this happen?

IF vLLM v0.15.0 Docker container on Blackwell GPUs with NVIDIA driver 590.48 and CUDA 13.0 crashes with 'Error 803: system has unsupported display driver / cuda driver combination' after the hang is fixed.

How others solved it

THEN Adjust `LD_LIBRARY_PATH` inside the container to ensure the host's NVIDIA driver libraries take precedence over the container's bundled CUDA libraries. For example, set `LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH` or remove `/usr/local/cuda/lib64` from the path. Alternatively revert to vLLM 0.14.1 until a backported fix (PR #33116) is available in v0.15.x.

In docker-compose, add `LD_LIBRARY_PATH: '/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}'` to the environment section.

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics